$ today_in_10_seconds
Anthropic quietly extended agentic tool use past the terminal and into physical equipment - microscopes, robotic arms, liquid handlers, quantum laser rigs. Meanwhile S&P Global is selling data retrieval to agents rather than analysts, Genpact hired a product chief to stop selling hours, and Amazon is closing the human microtask market that trained modern AI. Four different businesses re-pointing themselves at a non-human customer.
$ lead_story
Anthropic's Model Hardware Standard gives agents a driver layer for physical machines
On 27 August Anthropic opened a research preview of the Model Hardware Standard, a shared specification for AI agents to safely operate physical devices, starting with a first group of scientific research labs and advanced manufacturers. It works on any device with a programmable interface and lets devices and agents communicate across networks, with the stated goal of round-the-clock autonomous lab and manufacturing workflows running with minimal human intervention. Critically for anyone building agent harnesses, it's model-agnostic and reachable through MCP, CLI, or plain code and APIs - you are not buying into Claude to use it. One early result reported from the preview: QuEra's quantum laser stabilization went from 58% to 99.3%, and integration work that took weeks reportedly dropped to hours. It is not open source yet; Anthropic says it will open the spec after working through safety evaluations with early partners.
// why_it_matters: Every agent guardrail we've built assumes a failed action can be rolled back - and a mis-pipetted plate or a mis-aimed laser cannot.
$ also_worth_attention
Genpact hires a product chief to stop selling hours
Genpact appointed Priya Vijayarajendran as Chief Product and Platform Officer on 31 August, tasked with shifting the company toward productized, IP-led agentic solutions on a standardized enterprise platform. She arrives from ASAPP, where as CEO she consolidated point solutions into a single agentic platform running autonomous customer service for large enterprises, after leading Microsoft's Azure Data and AI business and serving as IBM's CTO for Cognitive, IoT and Watson Health. Read the CEO's line - we pivoted Genpact to be the Agentic Operations company - as what it is: a services firm admitting that billing for human process work is a shrinking business, and betting that owned process intelligence is the part agents can't commoditize.
link →S&P Global is selling retrieval to agents, not dashboards to analysts
Behind the data's-new-role think-pieces is a concrete product line. In July S&P launched Adaptive Retrieval, which lets customer agents and LLMs pull and assemble licensed S&P data from multiple sources via natural language, alongside a Deterministic option - two retrieval modes, one for tightly controlled pipelines and one for autonomous multi-agent work, both through its AI Data Portal. In August it extended the same data into Microsoft 365 Copilot with Excel connectors and agentic plugins. The strategic read: the incumbent data vendors have worked out that the buyer is now an agent, and they're competing on whether their data arrives cited, verifiable and auditable rather than on how good the terminal looks.
link →Amazon is closing Mechanical Turk on 30 September
After 21 years, the marketplace Bezos pitched as artificial artificial intelligence - which at its peak connected more than 500,000 workers to cent-scale tasks like labeling, transcription and moderation - shuts down at the end of the month, having stopped taking new customers on 30 July. The MTurk worker option disappears from SageMaker Ground Truth and Amazon Augmented AI on the same date, so check your pipelines. The detail that closes the loop: a 2023 study found as many as 46% of surveyed MTurk workers were already using AI to complete the tasks. The human-in-the-loop layer that trained the models is being retired by the models, and the specialist data platforms that replaced it price very differently.
link →
$ bigger_picture
The trend behind the headlines
Strip the announcements back and today is four organizations rewriting who their counterparty is. Anthropic made physical equipment a callable tool. S&P is selling retrieval to software rather than reports to humans. Genpact is trading billable process labor for IP that agents execute. Amazon is switching off the human task market that the whole field was built on. That's a coherent transition, and it puts unusual weight on this morning's point about reversibility: a wrong query costs a re-run, a wrong ledger entry costs a compensating transaction, and a wrong actuator command costs a broken instrument or a ruined sample. Anthropic's own caveat is the honest one to carry forward - physical and spatial reasoning still needs expert human oversight, which is another way of saying the preview is gated because the undo button doesn't exist.
PyLabRobot
Before MHS opens up, this is how you actually get an agent driving lab hardware today. It's a hardware-agnostic pure-Python library that puts one interface over Hamilton STAR and Vantage, Tecan Freedom EVO and Opentrons OT-2 machines, works on any modern OS, and ships a browser-based visualizer plus testing backends that need no physical robot attached. That simulator is the important part: it's where you let an agent fail before you give it an actuator.
$ subscribe --free
Get the next edition in your inbox
// DAIly Agent covers agentic AI three times a week
// free, no spam, exit anytime